Data-driven model analysis of the impact of environmental and socioeconomic factors on tuberculosis incidence

Overview

This study examines the determinants of tuberculosis incidence across 19 Taiwanese administrative divisions from 2014 to 2022 through predictive modeling. The research addresses TB epidemiology in the context of Taiwan's aging population and regional disease burden from Southeast Asian transmission corridors. The analysis integrates machine learning and deep learning methodologies to identify environmental and socioeconomic drivers while quantifying their relative contributions and interactions.

Methods and approach

The study deployed eight computational models comprising four machine learning algorithms (CatBoost, random forest, gradient boosting, and an unspecified fourth method) and four deep learning architectures across monthly TB incidence data from 19 Taiwanese jurisdictions. Twelve environmental and socioeconomic covariates were incorporated as predictive features. Model performance was evaluated and the three highest-performing approaches were selected for ensemble analysis. Explainable machine learning techniques were applied post-hoc to determine feature importance. Supplementary analysis incorporated stepwise regression and statistical assessment procedures to identify parsimonious configurations maintaining predictive accuracy with reduced dimensionality.

Key Findings

CatBoost, random forest, and gradient boosting models demonstrated superior predictive performance relative to other tested approaches. Six features emerged as primary determinants of TB incidence: population size, sulfur dioxide concentrations, physician density, normalized difference vegetation index, wind velocity, and precipitation. The analysis revealed nonlinear dose-response relationships and threshold effects between these variables and TB outcomes. Stepwise regression identified a reduced feature set capable of preserving high predictive accuracy.

Implications

The identification of population, atmospheric composition, healthcare infrastructure density, vegetation coverage, and meteorological variables as key TB drivers provides quantitative evidence for targeted public health intervention design. The demonstrated nonlinear relationships and threshold effects indicate that simple linear risk stratification models may underestimate complex epidemiological dynamics in TB transmission and progression. These findings establish a data-driven foundation for surveillance systems and resource allocation strategies in Taiwan. The methodological framework and variable set may be generalizable to TB forecasting and control efforts in other endemic regions with comparable epidemiological and environmental profiles.